robust-futex-ABI.txt 8.7 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182
  1. Started by Paul Jackson <pj@sgi.com>
  2. The robust futex ABI
  3. --------------------
  4. Robust_futexes provide a mechanism that is used in addition to normal
  5. futexes, for kernel assist of cleanup of held locks on task exit.
  6. The interesting data as to what futexes a thread is holding is kept on a
  7. linked list in user space, where it can be updated efficiently as locks
  8. are taken and dropped, without kernel intervention. The only additional
  9. kernel intervention required for robust_futexes above and beyond what is
  10. required for futexes is:
  11. 1) a one time call, per thread, to tell the kernel where its list of
  12. held robust_futexes begins, and
  13. 2) internal kernel code at exit, to handle any listed locks held
  14. by the exiting thread.
  15. The existing normal futexes already provide a "Fast Userspace Locking"
  16. mechanism, which handles uncontested locking without needing a system
  17. call, and handles contested locking by maintaining a list of waiting
  18. threads in the kernel. Options on the sys_futex(2) system call support
  19. waiting on a particular futex, and waking up the next waiter on a
  20. particular futex.
  21. For robust_futexes to work, the user code (typically in a library such
  22. as glibc linked with the application) has to manage and place the
  23. necessary list elements exactly as the kernel expects them. If it fails
  24. to do so, then improperly listed locks will not be cleaned up on exit,
  25. probably causing deadlock or other such failure of the other threads
  26. waiting on the same locks.
  27. A thread that anticipates possibly using robust_futexes should first
  28. issue the system call:
  29. asmlinkage long
  30. sys_set_robust_list(struct robust_list_head __user *head, size_t len);
  31. The pointer 'head' points to a structure in the threads address space
  32. consisting of three words. Each word is 32 bits on 32 bit arch's, or 64
  33. bits on 64 bit arch's, and local byte order. Each thread should have
  34. its own thread private 'head'.
  35. If a thread is running in 32 bit compatibility mode on a 64 native arch
  36. kernel, then it can actually have two such structures - one using 32 bit
  37. words for 32 bit compatibility mode, and one using 64 bit words for 64
  38. bit native mode. The kernel, if it is a 64 bit kernel supporting 32 bit
  39. compatibility mode, will attempt to process both lists on each task
  40. exit, if the corresponding sys_set_robust_list() call has been made to
  41. setup that list.
  42. The first word in the memory structure at 'head' contains a
  43. pointer to a single linked list of 'lock entries', one per lock,
  44. as described below. If the list is empty, the pointer will point
  45. to itself, 'head'. The last 'lock entry' points back to the 'head'.
  46. The second word, called 'offset', specifies the offset from the
  47. address of the associated 'lock entry', plus or minus, of what will
  48. be called the 'lock word', from that 'lock entry'. The 'lock word'
  49. is always a 32 bit word, unlike the other words above. The 'lock
  50. word' holds 3 flag bits in the upper 3 bits, and the thread id (TID)
  51. of the thread holding the lock in the bottom 29 bits. See further
  52. below for a description of the flag bits.
  53. The third word, called 'list_op_pending', contains transient copy of
  54. the address of the 'lock entry', during list insertion and removal,
  55. and is needed to correctly resolve races should a thread exit while
  56. in the middle of a locking or unlocking operation.
  57. Each 'lock entry' on the single linked list starting at 'head' consists
  58. of just a single word, pointing to the next 'lock entry', or back to
  59. 'head' if there are no more entries. In addition, nearby to each 'lock
  60. entry', at an offset from the 'lock entry' specified by the 'offset'
  61. word, is one 'lock word'.
  62. The 'lock word' is always 32 bits, and is intended to be the same 32 bit
  63. lock variable used by the futex mechanism, in conjunction with
  64. robust_futexes. The kernel will only be able to wakeup the next thread
  65. waiting for a lock on a threads exit if that next thread used the futex
  66. mechanism to register the address of that 'lock word' with the kernel.
  67. For each futex lock currently held by a thread, if it wants this
  68. robust_futex support for exit cleanup of that lock, it should have one
  69. 'lock entry' on this list, with its associated 'lock word' at the
  70. specified 'offset'. Should a thread die while holding any such locks,
  71. the kernel will walk this list, mark any such locks with a bit
  72. indicating their holder died, and wakeup the next thread waiting for
  73. that lock using the futex mechanism.
  74. When a thread has invoked the above system call to indicate it
  75. anticipates using robust_futexes, the kernel stores the passed in 'head'
  76. pointer for that task. The task may retrieve that value later on by
  77. using the system call:
  78. asmlinkage long
  79. sys_get_robust_list(int pid, struct robust_list_head __user **head_ptr,
  80. size_t __user *len_ptr);
  81. It is anticipated that threads will use robust_futexes embedded in
  82. larger, user level locking structures, one per lock. The kernel
  83. robust_futex mechanism doesn't care what else is in that structure, so
  84. long as the 'offset' to the 'lock word' is the same for all
  85. robust_futexes used by that thread. The thread should link those locks
  86. it currently holds using the 'lock entry' pointers. It may also have
  87. other links between the locks, such as the reverse side of a double
  88. linked list, but that doesn't matter to the kernel.
  89. By keeping its locks linked this way, on a list starting with a 'head'
  90. pointer known to the kernel, the kernel can provide to a thread the
  91. essential service available for robust_futexes, which is to help clean
  92. up locks held at the time of (a perhaps unexpectedly) exit.
  93. Actual locking and unlocking, during normal operations, is handled
  94. entirely by user level code in the contending threads, and by the
  95. existing futex mechanism to wait for, and wakeup, locks. The kernels
  96. only essential involvement in robust_futexes is to remember where the
  97. list 'head' is, and to walk the list on thread exit, handling locks
  98. still held by the departing thread, as described below.
  99. There may exist thousands of futex lock structures in a threads shared
  100. memory, on various data structures, at a given point in time. Only those
  101. lock structures for locks currently held by that thread should be on
  102. that thread's robust_futex linked lock list a given time.
  103. A given futex lock structure in a user shared memory region may be held
  104. at different times by any of the threads with access to that region. The
  105. thread currently holding such a lock, if any, is marked with the threads
  106. TID in the lower 29 bits of the 'lock word'.
  107. When adding or removing a lock from its list of held locks, in order for
  108. the kernel to correctly handle lock cleanup regardless of when the task
  109. exits (perhaps it gets an unexpected signal 9 in the middle of
  110. manipulating this list), the user code must observe the following
  111. protocol on 'lock entry' insertion and removal:
  112. On insertion:
  113. 1) set the 'list_op_pending' word to the address of the 'lock word'
  114. to be inserted,
  115. 2) acquire the futex lock,
  116. 3) add the lock entry, with its thread id (TID) in the bottom 29 bits
  117. of the 'lock word', to the linked list starting at 'head', and
  118. 4) clear the 'list_op_pending' word.
  119. On removal:
  120. 1) set the 'list_op_pending' word to the address of the 'lock word'
  121. to be removed,
  122. 2) remove the lock entry for this lock from the 'head' list,
  123. 2) release the futex lock, and
  124. 2) clear the 'lock_op_pending' word.
  125. On exit, the kernel will consider the address stored in
  126. 'list_op_pending' and the address of each 'lock word' found by walking
  127. the list starting at 'head'. For each such address, if the bottom 29
  128. bits of the 'lock word' at offset 'offset' from that address equals the
  129. exiting threads TID, then the kernel will do two things:
  130. 1) if bit 31 (0x80000000) is set in that word, then attempt a futex
  131. wakeup on that address, which will waken the next thread that has
  132. used to the futex mechanism to wait on that address, and
  133. 2) atomically set bit 30 (0x40000000) in the 'lock word'.
  134. In the above, bit 31 was set by futex waiters on that lock to indicate
  135. they were waiting, and bit 30 is set by the kernel to indicate that the
  136. lock owner died holding the lock.
  137. The kernel exit code will silently stop scanning the list further if at
  138. any point:
  139. 1) the 'head' pointer or an subsequent linked list pointer
  140. is not a valid address of a user space word
  141. 2) the calculated location of the 'lock word' (address plus
  142. 'offset') is not the valid address of a 32 bit user space
  143. word
  144. 3) if the list contains more than 1 million (subject to
  145. future kernel configuration changes) elements.
  146. When the kernel sees a list entry whose 'lock word' doesn't have the
  147. current threads TID in the lower 29 bits, it does nothing with that
  148. entry, and goes on to the next entry.
  149. Bit 29 (0x20000000) of the 'lock word' is reserved for future use.